[SimplifyLibCalls] Generalize replacePowWithSqrt to handle more cases.

[fhahn]

Summary

This patch folds the logic to expand

• pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x)))

into replacePowWithSqrt. If the original pow has NoInfs, we do not need
the select, if it has NoSignedZero, we do not need the fabs.

This addresses several FIXMEs in test/Transforms/InstCombine/pow-sqrt.ll.

One thing I am not entirely sure about: did we need isFast() in the original code for anything besides NoInfs and NoUnsignedZeros (e.g. errno behavior)?

[spatel]

Is this identical to D50036 / rL339887?

[fhahn]

Yep it is. What a coincidence :)

lib/Transforms/Utils/SimplifyLibCalls.cpp

static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
 
   InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
                                  getPow(InnerChain, AddChain[Exp][1], B));
   return InnerChain[Exp];
 }
 
 /// Use square root in place of pow(x, +/-0.5).
 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilder<> &B) {
-  // TODO: There is some subset of 'fast' under which these transforms should
-  // be allowed.
-  if (!Pow->isFast())
-    return nullptr;
-
   Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
   Type *Ty = Pow->getType();
 
   const APFloat *ExpoF;
   if (!match(Expo, m_APFloat(ExpoF)) ||
       (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
     return nullptr;
 
+  if (ExpoF->isExactlyValue(-0.5) &&
+      (!Pow->hasNoSignedZeros() || !Pow->hasNoInfs()))
+    return nullptr;
+
   // If errno is never set, then use the intrinsic for sqrt().
   if (Pow->hasFnAttr(Attribute::ReadNone)) {
     Function *SqrtFn = Intrinsic::getDeclaration(Pow->getModule(),
                                                  Intrinsic::sqrt, Ty);
     Sqrt = B.CreateCall(SqrtFn, Base);
   }
   // Otherwise, use the libcall for sqrt().
   else if (hasUnaryFloatFn(TLI, Ty, LibFunc_sqrt, LibFunc_sqrtf, LibFunc_sqrtl))
     // TODO: We also should check that the target can in fact lower the sqrt()
     // libcall. We currently have no way to ask this question, so we ask if
     // the target has a sqrt() libcall, which is not exactly the same.
     Sqrt = emitUnaryFloatFnCall(Base, TLI->getName(LibFunc_sqrt), B,
                                 Pow->getCalledFunction()->getAttributes());
   else
     return nullptr;
 
   // If the exponent is negative, then get the reciprocal.
   if (ExpoF->isNegative())
     Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
 
+  // Deal with x == -0. We expand pow(x, 0.5/-0.5) to fabs(sqrt(x)|1/sqrt(x)).
+  // This is faster than calling pow(), and still handles -0 correctly.
+  if (!Pow->hasNoSignedZeros()) {
+    Function *FAbsFn =
+        Intrinsic::getDeclaration(Pow->getModule(), Intrinsic::fabs, Ty);
+    Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
+  }
+
+  // Deal with x == -infinity. We expand
+  //  * pow(x, 0.5) to (x == -infinity ? +infinity : Sqrt))
+  // This is faster than calling pow(), and still handles negative infinity
+  // correctly.
+  if (!Pow->hasNoInfs()) {
+    Value *NegInf = ConstantFP::getInfinity(Ty, true);
+    Value *PosInf = ConstantFP::getInfinity(Ty);
+    Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
+    Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
+  }
   return Sqrt;
 }
 
 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilder<> &B) {
   Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
   Function *Callee = Pow->getCalledFunction();
   AttributeList Attrs = Callee->getAttributes();
   StringRef Name = Callee->getName();
Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilder<> &B) {
 
   // pow(x, 2.0) -> x * x
   if (match(Expo, m_SpecificFP(2.0)))
     return B.CreateFMul(Base, Base, "square");
 
   if (Value *Sqrt = replacePowWithSqrt(Pow, B))
     return Sqrt;
 
-  // FIXME: Correct the transforms and pull this into replacePowWithSqrt().
-  ConstantFP *ExpoC = dyn_cast<ConstantFP>(Expo);
-  if (ExpoC && ExpoC->isExactlyValue(0.5) &&
-      hasUnaryFloatFn(TLI, Ty, LibFunc_sqrt, LibFunc_sqrtf, LibFunc_sqrtl)) {
-    // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
-    // This is faster than calling pow(), and still handles -0.0 and
-    // negative infinity correctly.
-    // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
-    Value *PosInf = ConstantFP::getInfinity(Ty);
-    Value *NegInf = ConstantFP::getInfinity(Ty, true);
-
-    // TODO: As above, we should lower to the sqrt() intrinsic if the pow() is
-    // an intrinsic, to match errno semantics.
-    Value *Sqrt = emitUnaryFloatFnCall(Base, TLI->getName(LibFunc_sqrt),
-                                       B, Attrs);
-    Function *FAbsFn = Intrinsic::getDeclaration(Module, Intrinsic::fabs, Ty);
-    Value *FAbs = B.CreateCall(FAbsFn, Sqrt, "abs");
-    Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
-    Sqrt = B.CreateSelect(FCmp, PosInf, FAbs);
-    return Sqrt;
-  }
-
   // pow(x, n) -> x * x * x * ...
   const APFloat *ExpoF;
   if (Pow->isFast() && match(Expo, m_APFloat(ExpoF))) {
     // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
     APFloat LimF(ExpoF->getSemantics(), 33.0),
             ExpoA(abs(*ExpoF));
     if (ExpoA.isInteger() && ExpoA.compare(LimF) == APFloat::cmpLessThan) {
       // We will memoize intermediate products of the Addition Chain.

test/Transforms/InstCombine/pow-1.ll

 ;
   %r = call fast <2 x double> @llvm.pow.v2f64(<2 x double> %x, <2 x double> <double -1.0, double -1.0>)
   ret <2 x double> %r
 }
 
 declare double @llvm.pow.f64(double %Val, double %Power)
 define double @test_simplify17(double %x) {
 ; ANY-LABEL: @test_simplify17(
-; ANY-NEXT:    [[SQRT:%.*]] = call double @sqrt(double [[X:%.*]]) #2
+; ANY-NEXT:    [[SQRT:%.*]] = call double @llvm.sqrt.f64(double [[X:%.*]])
 ; ANY-NEXT:    [[ABS:%.*]] = call double @llvm.fabs.f64(double [[SQRT]])
 ; ANY-NEXT:    [[ISINF:%.*]] = fcmp oeq double [[X]], 0xFFF0000000000000
 ; ANY-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
 ; ANY-NEXT:    ret double [[TMP1]]
 ;
   %retval = call double @llvm.pow.f64(double %x, double 0.5)
   ret double %retval
 }

test/Transforms/InstCombine/pow-sqrt.ll

 ; CHECK-NEXT:    ret double [[TMP1]]
 ;
   %pow = call double @pow(double %x, double 5.0e-01)
   ret double %pow
 }
 
 define double @pow_intrinsic_half_no_FMF(double %x) {
 ; CHECK-LABEL: @pow_intrinsic_half_no_FMF(
-; CHECK-NEXT:    [[SQRT:%.*]] = call double @sqrt(double [[X:%.*]]) #1
-; CHECK-NEXT:    [[ABS:%.*]] = call double @llvm.fabs.f64(double [[SQRT]])
+; CHECK-NEXT:    [[TMP1:%.*]] = call double @llvm.sqrt.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[ABS:%.*]] = call double @llvm.fabs.f64(double [[TMP1]])
 ; CHECK-NEXT:    [[ISINF:%.*]] = fcmp oeq double [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
-; CHECK-NEXT:    ret double [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
+; CHECK-NEXT:    ret double [[TMP2]]
 ;
   %pow = call double @llvm.pow.f64(double %x, double 5.0e-01)
   ret double %pow
 }
 
 ; This makes no difference, but FMF are propagated.
 
 define double @pow_libcall_half_approx(double %x) {
 ; CHECK-LABEL: @pow_libcall_half_approx(
 ; CHECK-NEXT:    [[SQRT:%.*]] = call afn double @sqrt(double [[X:%.*]])
 ; CHECK-NEXT:    [[ABS:%.*]] = call afn double @llvm.fabs.f64(double [[SQRT]])
 ; CHECK-NEXT:    [[ISINF:%.*]] = fcmp afn oeq double [[X]], 0xFFF0000000000000
 ; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
 ; CHECK-NEXT:    ret double [[TMP1]]
 ;
   %pow = call afn double @pow(double %x, double 5.0e-01)
   ret double %pow
 }
 
-; FIXME
-
 define <2 x double> @pow_intrinsic_half_approx(<2 x double> %x) {
 ; CHECK-LABEL: @pow_intrinsic_half_approx(
-; CHECK-NEXT:    [[POW:%.*]] = call afn <2 x double> @llvm.pow.v2f64(<2 x double> [[X:%.*]], <2 x double> <double 5.000000e-01, double 5.000000e-01>)
-; CHECK-NEXT:    ret <2 x double> [[POW]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call afn <2 x double> @llvm.sqrt.v2f64(<2 x double> [[X:%.*]])
+; CHECK-NEXT:    [[ABS:%.*]] = call afn <2 x double> @llvm.fabs.v2f64(<2 x double> [[TMP1]])
+; CHECK-NEXT:    [[ISINF:%.*]] = fcmp afn oeq <2 x double> [[X]], <double 0xFFF0000000000000, double 0xFFF0000000000000>
+; CHECK-NEXT:    [[TMP2:%.*]] = select <2 x i1> [[ISINF]], <2 x double> <double 0x7FF0000000000000, double 0x7FF0000000000000>, <2 x double> [[ABS]]
+; CHECK-NEXT:    ret <2 x double> [[TMP2]]
 ;
   %pow = call afn <2 x double> @llvm.pow.v2f64(<2 x double> %x, <2 x double> <double 5.0e-01, double 5.0e-01>)
   ret <2 x double> %pow
 }
 
-; FIXME:
 ; If we can disregard INFs, no need for a select.
-
 define double @pow_libcall_half_ninf(double %x) {
 ; CHECK-LABEL: @pow_libcall_half_ninf(
 ; CHECK-NEXT:    [[SQRT:%.*]] = call ninf double @sqrt(double [[X:%.*]])
 ; CHECK-NEXT:    [[ABS:%.*]] = call ninf double @llvm.fabs.f64(double [[SQRT]])
-; CHECK-NEXT:    [[ISINF:%.*]] = fcmp ninf oeq double [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
-; CHECK-NEXT:    ret double [[TMP1]]
+; CHECK-NEXT:    ret double [[ABS]]
 ;
   %pow = call ninf double @pow(double %x, double 5.0e-01)
   ret double %pow
 }
 
-; FIXME:
-
 define <2 x double> @pow_intrinsic_half_ninf(<2 x double> %x) {
 ; CHECK-LABEL: @pow_intrinsic_half_ninf(
-; CHECK-NEXT:    [[POW:%.*]] = call ninf <2 x double> @llvm.pow.v2f64(<2 x double> [[X:%.*]], <2 x double> <double 5.000000e-01, double 5.000000e-01>)
-; CHECK-NEXT:    ret <2 x double> [[POW]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call ninf <2 x double> @llvm.sqrt.v2f64(<2 x double> [[X:%.*]])
+; CHECK-NEXT:    [[ABS:%.*]] = call ninf <2 x double> @llvm.fabs.v2f64(<2 x double> [[TMP1]])
+; CHECK-NEXT:    ret <2 x double> [[ABS]]
 ;
   %pow = call ninf <2 x double> @llvm.pow.v2f64(<2 x double> %x, <2 x double> <double 5.0e-01, double 5.0e-01>)
   ret <2 x double> %pow
 }
 
-; FIXME:
-; If we can disregard -0.0, no need for fabs.
-
 define double @pow_libcall_half_nsz(double %x) {
 ; CHECK-LABEL: @pow_libcall_half_nsz(
 ; CHECK-NEXT:    [[SQRT:%.*]] = call nsz double @sqrt(double [[X:%.*]])
-; CHECK-NEXT:    [[ABS:%.*]] = call nsz double @llvm.fabs.f64(double [[SQRT]])
 ; CHECK-NEXT:    [[ISINF:%.*]] = fcmp nsz oeq double [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
+; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[SQRT]]
 ; CHECK-NEXT:    ret double [[TMP1]]
 ;
   %pow = call nsz double @pow(double %x, double 5.0e-01)
   ret double %pow
 }
 
-; FIXME:
-
 define double @pow_intrinsic_half_nsz(double %x) {
 ; CHECK-LABEL: @pow_intrinsic_half_nsz(
-; CHECK-NEXT:    [[SQRT:%.*]] = call nsz double @sqrt(double [[X:%.*]]) #1
-; CHECK-NEXT:    [[ABS:%.*]] = call nsz double @llvm.fabs.f64(double [[SQRT]])
+; CHECK-NEXT:    [[TMP1:%.*]] = call nsz double @llvm.sqrt.f64(double [[X:%.*]])
 ; CHECK-NEXT:    [[ISINF:%.*]] = fcmp nsz oeq double [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
-; CHECK-NEXT:    ret double [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[TMP1]]
+; CHECK-NEXT:    ret double [[TMP2]]
 ;
   %pow = call nsz double @llvm.pow.f64(double %x, double 5.0e-01)
   ret double %pow
 }
 
-; FIXME:
-; This is just sqrt.
-
 define float @pow_libcall_half_ninf_nsz(float %x) {
 ; CHECK-LABEL: @pow_libcall_half_ninf_nsz(
 ; CHECK-NEXT:    [[SQRTF:%.*]] = call ninf nsz float @sqrtf(float [[X:%.*]])
-; CHECK-NEXT:    [[ABS:%.*]] = call ninf nsz float @llvm.fabs.f32(float [[SQRTF]])
-; CHECK-NEXT:    [[ISINF:%.*]] = fcmp ninf nsz oeq float [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], float 0x7FF0000000000000, float [[ABS]]
-; CHECK-NEXT:    ret float [[TMP1]]
+; CHECK-NEXT:    ret float [[SQRTF]]
 ;
   %pow = call ninf nsz float @powf(float %x, float 5.0e-01)
   ret float %pow
 }
 
-; FIXME:
-
 define double @pow_intrinsic_half_ninf_nsz(double %x) {
 ; CHECK-LABEL: @pow_intrinsic_half_ninf_nsz(
-; CHECK-NEXT:    [[SQRT:%.*]] = call ninf nsz double @sqrt(double [[X:%.*]]) #1
-; CHECK-NEXT:    [[ABS:%.*]] = call ninf nsz double @llvm.fabs.f64(double [[SQRT]])
-; CHECK-NEXT:    [[ISINF:%.*]] = fcmp ninf nsz oeq double [[X]], 0xFFF0000000000000
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[ISINF]], double 0x7FF0000000000000, double [[ABS]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call ninf nsz double @llvm.sqrt.f64(double [[X:%.*]])
 ; CHECK-NEXT:    ret double [[TMP1]]
 ;
   %pow = call ninf nsz double @llvm.pow.f64(double %x, double 5.0e-01)
   ret double %pow
 }
 
 ; Overspecified FMF to test propagation to the new op(s).
 

 ; CHECK-LABEL: @pow_intrinsic_neghalf_nsz(
 ; CHECK-NEXT:    [[POW:%.*]] = call nsz double @llvm.pow.f64(double [[X:%.*]], double -5.000000e-01)
 ; CHECK-NEXT:    ret double [[POW]]
 ;
   %pow = call nsz double @llvm.pow.f64(double %x, double -5.0e-01)
   ret double %pow
 }
 
-; FIXME:
 ; This is just recip-sqrt.
 
 define double @pow_intrinsic_neghalf_ninf_nsz(double %x) {
 ; CHECK-LABEL: @pow_intrinsic_neghalf_ninf_nsz(
-; CHECK-NEXT:    [[POW:%.*]] = call ninf nsz double @llvm.pow.f64(double [[X:%.*]], double -5.000000e-01)
-; CHECK-NEXT:    ret double [[POW]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call ninf nsz double @llvm.sqrt.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[RECIPROCAL:%.*]] = fdiv ninf nsz double 1.000000e+00, [[TMP1]]
+; CHECK-NEXT:    ret double [[RECIPROCAL]]
 ;
   %pow = call ninf nsz double @llvm.pow.f64(double %x, double -5.0e-01)
   ret double %pow
 }
 
-; FIXME:
-
 define float @pow_libcall_neghalf_ninf_nsz(float %x) {
 ; CHECK-LABEL: @pow_libcall_neghalf_ninf_nsz(
-; CHECK-NEXT:    [[POW:%.*]] = call ninf nsz float @powf(float [[X:%.*]], float -5.000000e-01)
-; CHECK-NEXT:    ret float [[POW]]
+; CHECK-NEXT:    [[SQRTF:%.*]] = call ninf nsz float @sqrtf(float [[X:%.*]])
+; CHECK-NEXT:    [[RECIPROCAL:%.*]] = fdiv ninf nsz float 1.000000e+00, [[SQRTF]]
+; CHECK-NEXT:    ret float [[RECIPROCAL]]
 ;
   %pow = call ninf nsz float @powf(float %x, float -5.0e-01)
   ret float %pow
 }
 
 ; Overspecified FMF to test propagation to the new op(s).
 
 define float @pow_libcall_neghalf_fast(float %x) {